mcp-interactive-instruction

MCP server for interactive instruction documents. Enables AI agents to autonomously manage documentation - creating drafts, organizing knowledge, and promoting to confirmed docs with user approval.

Why

Traditional approach of loading large .md files (like agents.md, skills.md) at the start of a conversation has limitations:

Context bloat: All documentation occupies context space even when not needed
Forgetting: As conversation grows, AI gradually "forgets" earlier loaded content
All or nothing: No way to selectively refresh specific information
Static knowledge: AI can't record new learnings from the conversation

This tool solves these problems by:

Topic-based splitting: Organize documentation into separate files by topic
On-demand retrieval: Fetch only what's needed, when it's needed
Interactive recall: AI can "remember" information by querying the MCP tool
Autonomous learning: AI can record new knowledge as drafts without permission
Human oversight: Drafts require approval before becoming confirmed docs

Tools

| Tool | Purpose | Permission | |------|---------|------------| | description | Show usage instructions for all tools | - | | help | Browse/read confirmed documentation | - | | draft | CRUD for temporary docs (_mcp_drafts/) | AI can use freely | | apply | Promote drafts to confirmed docs | Requires user approval |

description

Show detailed usage instructions for all MCP tools.

description()
→ Full usage guide with examples

help

Browse and read confirmed documentation. Drafts (_mcp_drafts/) are automatically filtered out.

# List root level (shows categories and documents)
help()

# Navigate into a category
help({ id: "git" })

# Get specific document content
help({ id: "git__workflow" })

# List ALL documents at once (flat view)
help({ recursive: true })

draft

Manage temporary documentation drafts. AI should use this freely to record any new information learned from user instructions.

# Show draft tool help
draft()

# List all drafts
draft({ action: "list" })

# Read a draft
draft({ action: "read", id: "coding__style" })

# Create new draft (NEW topic = NEW file!)
draft({ action: "add", id: "coding__testing", content: "# Testing Rules\n\n..." })

# Update existing draft (same topic only)
draft({ action: "update", id: "coding__testing", content: "# Testing Rules\n\nUpdated..." })

# Delete a draft
draft({ action: "delete", id: "old-draft" })

# Rename/move a draft (safe reorganization)
draft({ action: "rename", id: "old-name", newId: "category__new-name" })

Important Rules for AI:

New information = New file: Different topic = always use add, not update
One topic per file: Keep each draft focused on a single topic
Use hierarchy: Group related topics with prefixes (e.g., coding__testing)

apply

Promote drafts to confirmed documentation. This requires user approval.

# Show apply tool help
apply()

# List drafts ready to promote
apply({ action: "list" })

# Promote a draft (same name)
apply({ action: "promote", draftId: "coding-style" })
→ Moves _mcp_drafts/coding-style.md to coding-style.md

# Promote with different name/location
apply({ action: "promote", draftId: "temp-guide", targetId: "guides__setup" })
→ Moves _mcp_drafts/temp-guide.md to guides/setup.md

Installation

npm install -g mcp-interactive-instruction

Or use directly with npx:

npx mcp-interactive-instruction /path/to/docs

Configuration

Claude Code

Add to ~/.claude/settings.json for global configuration:

{
  "mcpServers": {
    "docs": {
      "command": "npx",
      "args": ["-y", "mcp-interactive-instruction", "/path/to/your/docs"]
    }
  }
}

Or create .mcp.json in your project root:

{
  "mcpServers": {
    "docs": {
      "command": "npx",
      "args": ["-y", "mcp-interactive-instruction", "./docs"]
    }
  }
}

Reminder Flags (Optional)

Optionally add flags to help AI remember to use the MCP tools:

{
  "mcpServers": {
    "docs": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-interactive-instruction",
        "./docs",
        "--remind-mcp",
        "--remind-organize",
        "--reminder", "Always check tests before committing"
      ]
    }
  }
}

| Flag | Effect | |------|--------| | --remind-mcp | Reminds AI to check docs before starting tasks | | --remind-organize | Reminds AI to keep docs organized (1 topic per file) | | --reminder <message> | Add custom reminder message (can be used multiple times) | | --topic-for-every-task <id> | Specify a document AI must re-read before every task | | --info-expires <seconds> | How long MCP info stays valid (default: 60). Works with --topic-for-every-task |

Topic for Every Task

Force AI to re-read a specific document before every task. This is useful for critical rules that should never be forgotten:

{
  "args": [
    "-y",
    "mcp-interactive-instruction",
    "./docs",
    "--topic-for-every-task", "topic-for-every-task",
    "--info-expires", "60"
  ]
}

The --info-expires flag tells AI that MCP information expires after N seconds and needs to be refreshed. This triggers re-reading the specified document before each task.

Best Practice: Keep the topic-for-every-task document as a redirect hub rather than a detailed rule list:

# Topic for Every Task

Read these documents before starting any task:

- `why-this-project` - Project concept and goals
- `coding-rules` - Essential coding conventions

## Quick Reminders
- Use params object style for function arguments
- All documentation must be in English

This approach keeps the document lightweight while ensuring AI always knows which topics to check.

Tuning --info-expires: Shorter expiry times cause more frequent re-reads, ensuring rules are never forgotten. However, this consumes more context space. Adjust based on your needs:

| Value | Effect | |-------|--------| | 30-60s | Frequent re-reads, higher context usage | | 120-300s | Balanced approach | | 600s+ | Rare re-reads, lower context usage |

Note: This feature influences AI behavior but does not guarantee 100% compliance. AI may still make autonomous decisions about when to re-read documents based on context and task requirements.

Example with multiple custom reminders:

{
  "args": [
    "-y",
    "mcp-interactive-instruction",
    "./docs",
    "--reminder", "Run tests after code changes",
    "--reminder", "Use Japanese for commit messages"
  ]
}

Directory Structure

docs/
├── coding-style.md              → id: "coding-style" (confirmed)
├── git/
│   ├── workflow.md              → id: "git__workflow" (confirmed)
│   └── commands.md              → id: "git__commands" (confirmed)
└── _mcp_drafts/                 ← AI's temporary drafts
    ├── new-feature.md           → draft id: "new-feature"
    └── coding/
        └── testing.md           → draft id: "coding__testing"

Confirmed docs: Root level and subdirectories (excluding _mcp_drafts/)
Drafts: Stored in _mcp_drafts/ directory
ID Format: Use __ (double underscore) as the path separator

Workflow

For AI

Check docs before tasks: Use help() to see available documentation
Record new learnings: When user teaches something new, immediately create a draft
One topic per file: Keep drafts focused and granular
Ask before promoting: Get user approval before using apply

# User says: "Always use params object style for function arguments"
draft({ action: "add", id: "coding__params-style", content: "# Params Style\n\n..." })

# Later, ask user: "Should I promote this to confirmed docs?"
apply({ action: "promote", draftId: "coding__params-style" })

For Users

Review drafts: Check draft({ action: "list" }) to see what AI has recorded
Approve or reject: Decide which drafts should become permanent documentation
Organize: Use apply with targetId to place docs in the right location

Document Format

# Title

Summary paragraph that appears in the document list.

## Section 1

Content...

The first paragraph after the # Title heading is used as the summary in listings. Make summaries descriptive so AI can identify when each document is relevant.

Granularity Guidelines

Keep each document focused on ONE topic:

| Instead of | Split into | |------------|------------| | git.md (everything) | git__workflow.md + git__commands.md | | coding.md (all rules) | coding__style.md + coding__testing.md |

Why this matters:

AI loads only what's needed
Easier to find and update specific information
Better summaries for matching

Performance

Caching: Document list cached for 1 minute
Cache invalidation: Automatic on write operations
Lazy loading: Documents read only when requested

License

MIT

MCP Servers